Geocoding accuracy: match rates, match types, and handling failures

Reference article · 7 min read

A service request that shows up in the middle of a river, a household assigned to the wrong council district, a heat map with a bright spot at the exact center of a ZIP code: these are usually geocoding artifacts, not data entry errors. Geocoding turns a text address into a coordinate, and every coordinate carries uncertainty that the output rarely displays. When a map is visibly wrong, the first job is not to fix it but to quantify it, because the size and shape of the error determine whether the map is unusable, usable with caveats, or fine for the question being asked.

Three fields describe geocoding quality: how many records got a coordinate, how that coordinate was derived, and whether the answer was unambiguous. Agencies that store all three alongside the coordinate can defend their maps. Agencies that store only latitude and longitude cannot, because the record no longer remembers how much to trust itself.

What a match rate actually tells you

Match rate is the share of input records that received a coordinate. It is a coverage measure, not an accuracy measure. A geocoder can push its match rate close to complete by falling back to progressively coarser geographies, returning a city centroid rather than admitting failure. Reported alone, a high match rate can conceal a large population of records placed at the center of a polygon miles wide.

Track match rate broken out by match type, and track it by area as well as in aggregate. Failures cluster: new subdivisions, recently annexed territory, rural route addresses, unit numbers in large complexes, and intersections or landmarks used in place of addresses. An overall figure that looks acceptable can hide a neighborhood where most records failed, which is exactly the neighborhood where the map will be challenged.

Match types, from address point to centroid

Terminology varies by product, so require a written crosswalk from any vendor. The common tiers, from most to least precise:

• Address point or rooftop. The coordinate comes from a maintained point feature representing a specific address, often near the structure or its access point. Quality depends entirely on who maintains the point layer and how current it is.
• Parcel centroid. The geometric center of the parcel polygon. Usually good, but on large or irregular parcels the center can sit far from the building, and multiple addresses on one parcel collapse to the same point.
• Street interpolation. The geocoder estimates a position along a centerline segment using the address range assigned to that segment. Error grows with segment length, uneven lot sizes, vacant land, and address numbers outside the stored range. Most implementations offset the point to one side of the line.
• Street segment or block centroid. The midpoint of a block, returned when the specific number cannot be placed.
• ZIP, place, or county centroid. A single point standing in for a whole delivery or administrative area. Useful for rough regional counts, misleading on any map that shows individual incidents.

Ties, confidence scores, and silent substitutions

A tie occurs when more than one candidate matches equally well: the same number on a north and south variant of a street, the same street name in two municipalities within a county, or a record missing a directional prefix. The default behavior matters more than the tie rate. A geocoder that returns all candidates lets staff resolve the ambiguity. One that silently picks the first candidate produces confident, wrong coordinates that no one reviews.

Confidence scores are vendor-specific and are not distances. A score of 90 in one product is not comparable to 90 in another, and neither converts to meters. Treat scores as a sorting tool for review queues, not as an accuracy claim. Also capture the standardized address the geocoder actually matched, because silent substitutions, a dropped unit number or a guessed suffix, are a common source of plausible but incorrect results.

Match precision to the decision being made

• Boundary assignment. District, precinct, attendance zone, service area, and taxing jurisdiction assignments need address point or parcel precision. Jurisdiction boundaries frequently follow street centerlines, and an interpolated point sitting on or near the centerline can land on either side.
• Dispatch and field response. Requires verified address points, and often supplemental access information that a coordinate alone does not carry.
• Density and hot spot analysis. Block level is often sufficient, but centroid fallbacks create artificial clusters. Exclude coarse match types from the analysis rather than letting them accumulate at a single pixel.
• Program or grant eligibility tied to a geography. Tract, zone, and boundary based eligibility determinations should be documented and reproducible. Program rules and the geographies they reference change, so confirm the current requirement with the administering agency before relying on a stored result.

What to do with records that do not match

Dropping unmatched records silently biases every number downstream, and the bias is rarely random. Keep them and give them a workflow.

1. Retain the original input text, the match status, match type, score, reference source, and the date and version of the reference data used.
2. Triage failures by cause: standardization problems, missing reference features, non-address locations such as intersections and mile markers, multi-unit addresses, and post office boxes.
3. Route corrections to the authoritative address or parcel source so the fix persists, rather than editing the analysis copy.
4. Maintain a manual review queue with written placement rules, and record who placed each manual point and when.
5. Publish coverage with every map and report: how many records were geocoded, at what precision, and how many were not placed at all.

Measuring quality on your own data

Vendor accuracy claims are measured on someone else's addresses. Build a local truth set from authoritative address points or parcels, sampled across dense urban, suburban, rural, new construction, and multi-unit conditions. Then measure positional difference against the truth point, the share of records that land in the same boundary as the truth point, and the false positive rate, meaning records that matched confidently to the wrong location. Test with real intake text, including abbreviations and typos, not a cleaned list. Re-run the same test after each reference data update and keep the harness.

What to ask vendors and require in writing

• Which reference datasets are used, how often they are refreshed, and whether the service can consume your local address points and parcel layer as the primary source.
• Whether each response returns match type, score, tie candidates, the standardized address, and an identifier for the reference feature used.
• The exact definition behind any "rooftop" label, since it sometimes means a parcel centroid.
• Whether results are reproducible, including the ability to pin or record a reference data version.
• How submitted addresses are handled: retention period, logging, and whether inputs are used for any purpose beyond returning a result.
• Deployment options, batch limits, and throughput for periodic reprocessing of historical records.

In the contract, require the per-record metadata fields as a schema, a documented match type crosswalk, acceptance testing against your own sample with thresholds you set, notice before reference data changes, and full export of results at termination. Addresses tied to individuals may be sensitive under state public records and privacy law, and treatment differs by record type and jurisdiction, so confirm handling and any restrictions on sending addresses to third-party services with your counsel and records officer.

HQT works with state and local agencies on data and analytics engineering, including address data quality and geocoding pipelines.

Frequently asked questions